Cause And Effect In Epidemiology Transcript - NWCPHP

Cause and Effect in EpidemiologyTranscriptprocess. For instance, we can talk about smoking beingC a u sa l A sso cia tio n in E p id e m iolog ya cause of lung cancer even though some people whohave never smoked also get lung cancer—smoking is not aE p id e m io lo gists u se a d efin itio n of cause th a t is d iffe ren tnecessary factor for all cases of lung cancer.from o the r d iscip lin e s. C ause is a factor that plays a role in producing anIn the most general sense, a cause is something that if itoccurrence of the disease.weren’t there, some cases of the disease wouldn’t happen. T he causal factor is not a n ecessary part of the dise aseprocess.This definition allows that factors can play a direct role C ause is som ething that if it w eren’t there, som e cases ofor an indirect role in causing disease. A factor may notthe disease w ouldn’t happen.be capable of causing disease all by itself; it may be just T he causal factor can pla y a direct or ind irect role in ca usingdiseas e.one part of a more complex mechanism that necessar C ausality is not proven in any one study.ily involves other exposures or factors. For instance, notall smokers get lung cancer—smoking is not sufficient allby itself to cause lung cancer in all smokers. But we stillconsider smoking to be a cause of lung cancer.The key feature of the notion of cause and causality is that causality is notproven in any one study. It’s a process of determination or decision-makingor inference based on a variety of information, as we’ll discuss for the rest ofthis module.Causality TermsC a u sa lity T e rm sLet’s talk for a moment about some terms with specificmeanings to epidemiologists.A sso cia tio n s a re observed. C a u se s a re inferred.Again, as a reminder, we observe associations—they areO b se rve d po sitive a sso ciatio nIn fe re n ce of ca usa tion S m ok ing increases risk of lun g cancer.the results of specific studies. And we infer causes through S m ok ing is a risk factor for lun g cancer.a process of decision-making that often uses the guidelinesO b se rve d neg a tive associa tionIn fe re n ce of p ro te ctionwe’ll cover later in this module. E xercise decreases risk of heart d isease.An observed positive association, such as between E xercise protects against heart d isease.smoking and lung cancer, could lead us to an inferenceof causation. We would then say that smoking increasesrisk of lung cancer, that is, smoking is a risk factor for lungcancer.An observed negative association, such as that between exercise and heartdisease, could lead us to an inference of protection. We would then say thatexercise decreases risk of heart disease, or protects against heart disease.These statements, specifically the use of the words risk factor and protective factor, imply that you have made a decision about the causal nature ofthe relationships between the exposures and the outcomes under study.Now we will pause for the first of several interactive exercises about thematerial we have just covered. Please note that the exercises sometimes takeseveral seconds to load.Exercise 1Northwest Center for Public Health Practice4

Cause and Effect in EpidemiologyTranscriptCausal Inference GuidelinesC a u sa l In fe re n ce G u id e lin e sNow lets talk about guidelines epidemiologists use forcausal inference.E sse n tia l th a t a n a ssocia tio n is o b se rve d .First, and foremost—it’s essential that an associationC a u sa lity is n o t p ro ve n b y a n y o n e stu d y.be observed in order to proceed along the path of deter1. R and om ized tria l evidence exists.exposure2. N o a lternative exp la natio ns exist.mining whether there’s a cause-and-effect relationship.3.Timingoftherelationshipiscorrect.So let’s say we observe an association between a certain4. A ssociatio n is strong.exposure and a certain disease, and we want to know if5. A ssociatio n is b io lo gic ally p laus ib le.that exposure truly is a cause of that disease.6. H igher expos ures le ad to h igher risk s.Since causality is not proven in any one study, how do7. O bserved e vide nce is cons istent.we determine if an exposure causes a disease? This is animportant decision for public health practitioners to beable to make, as it may be the basis for determining whether to mount acampaign to decrease this exposure in a community.This list of guidelines may help structure your thinking about the meaningof observed associations, to help you decide whether you can infer causality in specific situations. In the rest of the module we will discuss these sevenguidelines:1. Randomized trial evidence exists2. No alternative explanations exist (or, as epidemiologists say, there is noconfounding)3. The timing of the relationship is correct (that is, the exposure comesbefore the disease)4. The association is strong5. The association is biologically plausible (that is, we know what themechanism might be)6. Higher doses of the exposure lead to progressively higher disease risk7. And finally, the observed evidence of the association is consistent.Let’s consider the first of these features: randomized trial evidence.1. Randomized Trial Evidence ExistsThe findings of randomized studies provide the strongestevidence pointing toward causality, because in these studies chance alone dictates which participants are exposedand which are unexposed.In randomized trials a group of people is assigned toreceive an exposure or an intervention, and these peopleare then followed over time to determine what proportionof them develop the target outcome under study, whichcould be an illness but could also be a beneficial outcomesuch as a decrease in blood pressure. At the same time, agroup of people is assigned not to receive the exposure,Northwest Center for Public Health Practice?diseas e1 . R a n d o m ize d T ria l E vid e n ce E xistsC h a n ce a lo ne d icta te s w h ichp a rticip a n ts of th e study a ree xp o se d .O utcom eE xposedN o outcom eO th e r fa cto rs d on ’t d isto rtth e re su lts.C a n ’t fe a sib ly stu d y a llqu e stio n s of ca u sa tio nw ith ra n d om ize d tria ls.N o t e th ica l to u se ran do m ize dstu d ie s fo r som e typ e s of riskfa cto rs.O utcom eN ot expose dN o outcom e R e ly o n observation a l stud ies.5

Cause and Effect in EpidemiologyTranscriptand this group is also followed up to determine how many develop theoutcome, and whether that proportion differs from the proportion in theexposed group.Because the researcher, rather than the participant, decides who will bein which group, other factors that could influence the risk of disease or thehealth event generally will not distort the results.Unfortunately, we cannot feasibly study all questions of causation withrandomized trials, and it isn’t ethical to use a randomized trial to study somerisk factors, such as suspected carcinogens. For these types of questions wemust rely on observational studies such as case-control and cohort studies.See the module on Study Types in Epidemiology for more information ontypes of observational study designs.Randomized Trial ExampleR a n d o m ize d T ria l E xa m p leHere’s an example of a randomized trial that has hadDoes supplementation with folic acid (FA) or with other vitaminswide-ranging public health effects. The purpose of the trialprevent neural tube defects (NTD)?F o lic acidwas to determine whether supplementation with folic acidM ed ica l R ese arch C ou ncil V itam inN TD b irth s298Studyor a mixture of other vitamins around the time of concep 33 centers in 7 countriesF o lic acid2 4 6tion could prevent neural tube defects (which are serious& vita m in s A ll partic ipa nts had pre vio u schildren w ith N T D295birth defects that include spina bifida and anencephaly).R esu ltsN o su pp le This study—called the medical research council vitamin A n y folic ac id: 6/593 w ith N T Dm e n ts N o folic ac id: 2 1/60 2 w ith N T D300study—was conducted in seven European countries in the13 8 21Vitam in s“ folic acid su pp lem entatio nlate 1980s. All of the women enrolled in the trial had hadstarting b efore pregnanc y can302no w b e firm ly recom m ended ”a previous child with a neural tube defect and thus theyLancet 1991; 338:131-7were at high risk of another pregnancy complicated by thisproblem.The women were divided into four groups and assigned to take differentcombinations of folic acid and other vitamins just before and during pregnancy. One group took folic acid supplements only, one group took folic acidplus other vitamins, one group took no supplements, and the final group tookonly the other vitamins.The results of this study were striking. The nearly 600 women who werein the groups assigned to take folic acid (whether with or without othervitamins) had only six children with neural tube defects. The similar-sizedgroup of women assigned to take other vitamins alone or no vitamins had 21affected children. The relative risk was 3.5, and this difference was statistically significant. Thus, we would say there was a strong protective effect offolic acid supplementation on the basis of this study.The authors concluded that “folic acid supplementation starting beforepregnancy can now be firmly recommended for all women who have hadan affected pregnancy, and public health measures should be taken toensure that the diet of all women who may bear children contains an6adequate amount of folic acid.”Northwest Center for Public Health Practice

Cause and Effect in EpidemiologyTranscriptAs a result of this and a few other similar studies, folic acid supplementation is now routine in pregnancy and some foods likely to be eaten bywomen of childbearing age are fortified with folic acid. You can see howpersuasive a randomized trial can be in determining a cause and effectassociation.2. No Alternative Explanations Exist2 . N o A lte rn a tive E xp la n a tio n s E xistOur next guideline applies particularly in studies that arenot randomized trials. We need to conclude that no alterThe increase in disease risk with a certain factor may not bedue to the action of that factor (an alternative explanationnative explanation exists for the association seen.could exist).It’s important to consider whether the increase inConfounding: A m ixin g o f effe cts. A ssociatio n seen b etw een exposure an d d isease is a distortiondisease risk we see in the presence of a certain factor isExposureDiseasedue instead to other co-occurring factors or exposures.We call this situation confounding. Before we can inferFactora causal relationship, we first must consider the possibilF o r co nfo u n d in g to o ccu r th e re m u st b e a n a sso cia tio nb e tw e e n :ity of confounding and either dismiss it or take it into the exposure of interest an d the extrane ous factor the extrane ous factor and the d isease of interestaccount. We can take confounding into account at eitherthe design or analysis stages of a study.What is confounding, exactly? It’s a mixing of effects. If there is confounding, the association that we see between an exposure and a disease is adistortion. This distortion occurs because another factor or exposure thathappens along with the one we’re interested in is also associated with thedisease we’re studying—so what we’re really seeing is a mixture of theeffects of two exposures or factors on the disease.In order for confounding to occur two things need to happen: First, thereneeds to be an association between the exposure you’re interested in andthis extraneous, or confounding, factor. And second, there needs to be anassociation between the extraneous factor and the disease you’re studying.Let’s use an example to illustrate this principle.Confounding ExampleHere we see the results of a study on whether alcoholdrinking is a risk factor for lung cancer, that is, are peoplewho drink alcohol more likely to get lung cancer thanthose who don’t? This study also looked at whether, amongthose who drink, increasing amounts of alcohol consumedin a day leads to increasing cancer likelihood.We see the crude, or unadjusted, results of the studyhere. It appears that alcohol intake is positively associatedwith lung cancer. We see that drinkers are more likely toget lung cancer, and those who drink the most heavily arethe most likely to get it.Northwest Center for Public Health PracticeC o n fo u n d in g E xa m p leA lco h o l co n su m ptio n an d th e risk of lu n g ca nce rA lcohol intake(g ram s/day)Lung cancerIncidence/10,000C rudeRRA djustedRR 07.41.01.00.1–1 213.61.81.012.1– 2416.42.21.0 2425.23.41.1Is there confounding by sm ok ing ? P eo p le w ho drink , m ore lik ely to sm ok e S m ok ers m ore lik ely to g et lung cancer adjusted for p ack-years of sm oking a nd oth er factorsS ee N W C P H P m odule on m ea sure s of associa tion for the d efinition a nd calculation o f a relativ e risk .7

Cause and Effect in EpidemiologyTranscriptBut before we decide that this is a causal association however, we haveto ask whether confounding is responsible for these associations. You mightspecifically wonder if there is confounding by smoking. Why? Becausesmoking meets our first requirement for a confounder. An association existsbetween the exposure, alcohol drinking, and the extraneous factor, smoking. Research has shown that people who drink alcohol are more likely tosmoke than people who don’t drink. So we may have a mixing of the effectsof alcohol and smoking when we think we’re looking only at the effects ofalcohol use.And why might that matter? Because smoking also meets our other requirement for confounding—there’s an association between smoking and thedisease we’re looking at, here, lung cancer. Smokers are more likely toget lung cancer than nonsmokers. So, the possibility exists that smoking is confounding the association between alcohol use and lung cancer.The effect we think is due to alcohol use may be due instead to smokingbecause many drinkers also smoke. To check this we conduct analyses that“adjust” for smoking, and when we do so, we see that there is no associationbetween alcohol use and lung cancer. Drinkers are no more likely to get lungcancer than nondrinkers. Controlling or adjusting for the confounder, smoking, has removed the association between alcohol use and lung cancer, andtherefore the association cannot be causal.3. Timing of Relationship Is Correct3 . T im in g o f R e latio n sh ip Is C o rre ctAn essential feature of an association in order to beT he su sp e cted cau se m u st com e b efo re th e d ise a se .considered as causal is that the timing is correct. The E xposure-outcom e tim e sequence ca n be difficult to e stablish.suspected cause must come before the disease. Not only In p rosp ective coh o rt stud ies or ra n d om ize d trials e xp o su res n o te da t b eg inn ing o f stud y.must the cause or exposure come before the effect or In case -co ntro l o r re tro sp e ctive coh o rt stud ies pa st e xp o su resco nside re d .disease but there must be enough time for the suspected C an b e d ifficult to determ ine tim e sequence w hen tim e perio dcause to have an effect.betw e en suspecte d cause and effect is short.The exposure-outcome time sequence can be difficultgivenAspirinforgivenfluto establish, and it’s easier in some type of studies thanReye’sorforchickenflu or2-3 daysSyndromechickenAspirinpoxin others. In prospective cohort studies and randomizedtim etrials, exposures are noted at the beginning of the study,at a time when the study participants are determined tobe free of disease, so it’s easy to determine the time sequence. In contrast,in case-control or retrospective cohort studies records of exposures areobtained or subjects are interviewed about past exposures. In these studiesit’s not always clear whether the exposure of interest occurred before thedisease process began.It may also be difficult to determine the time sequence when the timeperiod between the suspected cause and the effect is short. For example,Reye’s Syndrome, which is a serious neurological disease, had an appar-S ee N W C P H P m odule on stud y design s for m ore info rm ation on typ es o f epid em iologic studies .Northwest Center for Public Health Practice8

Cause and Effect in EpidemiologyTranscriptent sharp increase in occurrence in children in 1980, and aspirin use wassuspected as a cause. Researchers hypothesized that giving a child aspirinfor flu or chicken pox increased the risk of the disease in the next few days.One concern raised early on about the proposed association was that children in the early stages of Reye’s syndrome may have been given aspirin as atreatment for the disease, and thus one shouldn’t conclude that any associations seen were cause-and-effect because the effect, Reye’s Syndrome, mayhave come before the suspected cause, taking aspirin. Subsequent studies ofthis issue clarified that this wasn’t the case.Timing ExampleT im in g E xa m p leI’d like now to introduce an exposure-disease questionthat is of current public health importance. Does maternalD oes m aternal sm ok ing duringpreg nancy result in low er infantsmoking during pregnancy result in, or cause, lower birthbirthw eig ht?P rospective cohort study: askweights? To answer this question we might first considerw om en about sm ok ing beforethe idea of a randomized trial for a definite answer.and during preg nancy, thenm easure infant birthw eig ht atHowever, in this case a randomized trial is unethical anddelivery.R etrospective studyimpractical. Obviously we cannot tell some women toS w iss infants born b etw e ensmoke during pregnancy!O ct. 1993 – S ep. 1 994So we turn instead to observational studies to addressthis question, and we must work a little harder to makecausal inferences. The ideal observational study would bea prospective cohort study—in which women are asked about their smoking habits before and during pregnancy (perhaps several times) and then theinfant is weighed at birth. This process ensures that the exposure (smoking)came before the effect (lower birthweight). Because of the availability ofbirth certificate data, studies of this question have often been retrospectivein nature however—with information entered about smoking during pregnancy only after the birth has occurred.One such study was done in Switzerland in the early 1990s. Using birthcertificates, this study found that 76% of women who delivered during a oneyear period had never smoked, and 4% of them had low birthweight infants.5% of the women were ex-smokers who had smoked only before that pregnancy—and only 3% of them had low birthweight infants. Finally, 19% of thewomen were listed as smoking during pregnancy, and 11% of these womenhad low birthweight infants.Because the effect of smoking on an infant’s birthweight will occur beforethe infant is born, it’s clear that the cause comes before the effect here, andthis guideline for inferring causation is met. The fact that no effect on birthweight was seen among former smokers is further evidence that the important time period is during the pregnancy.Let’s pause now while you consider some questions about what you’vejust learned.Northwest Center for Public Health Practicesmokingsmokinglow birthweighttim eduring pregnancy11%4%3%M othersneversm okedM otherssm oked beforepregnancyM others sm okedduringpregnancy9

Cause and Effect in EpidemiologyTranscriptExercise 24. The Association Is Strong4 . T h e A sso cia tion Is S tro n gNow let’s return to our list of guidelines. Keep in mind thatthis guideline and all the following guidelines apply only(T h is gu id e lin e a nd th e fo llo w in g o n e s a pp ly o n ly w h e ne xp o su re co m e s befo re d ise a se a nd th e re is n oto studies in which the exposure clearly comes before theco nfo un d in g.)T he la rge r th e va lu e of th e re la tive risk, th e le ssdisease and there is no confounding.like ly th e a sso ciatio n is to b e fa lse .Perhaps the most intuitive guideline for causal infer S treng th of association is not a biolog icallyconsistent feature; the size of the relative riskence is that of strength of the association. In evaluatingdepends on the prevalence of other causes.the strength of the association between the suspected T he relative risk does not have to belarg e to infer causality.cause and the effect, the larger the value of the relative N utrition a l associatio ns w ith dise aseusually sm all, but p atterns em ergerisk or the odds ratio (for a positive association), the lessind icatin g causa lity.S ee NW C P H P m odule on m easures of association for m orelikely the association is to be false.inform ation on the definition and calculation of a relativ e risk.Fo

What is Epidemiology in Public Health? Data Interpretation for Public Health Professionals Study Types in Epidemiology Measuring Risk in Epidemiology We introduce a number of new terms in this module. If you want to review their definitions at any time, the glos-sary in the attachments link at the top of the screen may be useful. Course Objectives